Cerebras WSE-3 Turbo Delivers 10x Token Capacity in CS-4 System

Cerebras released the WSE-3 Turbo chip and the accompanying CS-4 system on August 18, 2024. This launch targets enterprise AI workloads by delivering significantly higher token throughput for large language models. Buyers focused on inference speed gain a platform that claims to outperform traditional GPU clusters by a wide margin.

New chip doubles bandwidth and compute for large language models

The WSE-3 Turbo chip serves as the core of the new CS-4 system. It integrates 900,000 cores and provides 44GB of on-chip SRAM. The CS-4 system architecture scales by integrating three of these WSE-3 Turbo chips into a single unit.

Specifications

• Core Count: 900,000
• On-chip SRAM: 44GB
• System Configuration: 3x WSE-3 Turbo chips
• Token Throughput (GPT-OSS): 4465 Token/s
• Inter-chip Latency: 2μs

Cerebras states that the new hardware doubles key performance metrics compared to the previous WSE-3 generation. These doubled metrics include FP16 sparse AI compute, memory bandwidth, on-chip interconnect bandwidth, and I/O bandwidth. The system also features low-latency 2-microsecond inter-chip connections within the CS-4 configuration.

Performance benchmarks highlight the system's capacity for ultra-high parameter models. Cerebras reports a token throughput of 4465 tokens per second when running the OpenAI GPT-OSS model. The company claims this speed is 30 times faster than comparable GPU solutions. The system can also sustain over 1000 tokens per second at a 10 trillion parameter scale.

The CS-4 system offers 10 times the token capacity and 2 times the speed of the earlier CS-3 system. This hardware shift allows enterprises to handle larger model sizes with greater efficiency. The company highlights that the new Nexus rack platform features a 50% reduction in component count and improved power delivery.

Cerebras positions the WSE-3 Turbo and CS-4 as a direct upgrade for high-performance AI inference. The platform aims to reduce latency and increase throughput for massive language models. The release represents a significant advancement in hardware evolution, offering 10 times the token capacity and double the speed compared to the previous CS-3 generation.